Noise-immune CatBoost And Kalman Filter For Short-term Traffic Flow Forecasting

As a key technology of traffic guidance and traffic management,short-term traffic flow prediction is a very important part of intelligent transportation system.Accurate and timely traffic flow information directly affects the successful deployment of intelligent transportation system.However,due to the inherent randomness and external noise of the traffic flow,the traffic flow data presents strong uncertainty and non-linearity,which is often turned into up and down fluctuating trend and intermittent oscillation.The prediction model needs to distinguish which are noise and which are subtle clues reflecting the sudden change of the traffic flow.Therefore,it is still a challenging task to find a robust and accurate prediction algorithm.The research work of this paper mainly focuses on the impact of noise on short-term traffic flow prediction: the goal is to retain the abrupt details of original data to the greatest extent while filtering out noise,so as to more accurately reflect the real traffic flow conditions.In order to better distinguish between the noise in the raw data and mutations in details,based on the wavelet transform to the original noisy data preprocessing,and puts forward two features based on traffic flow data itself to respond effectively to the abnormal value prediction model,at last,by a benchmark data and 14 classic models,such as MAPE,RMSE performance evaluation as the basis,The comparative experiment proves the validity of the prediction model proposed in this paper.The main work is as follows:(1)Noise immune Cat Boost short-term traffic flow prediction model(Nicat Boost)with time series features: firstly,denoising and enhancing the original signal through wavelet threshold denoising;Then the model is trained based on Catboost.Compared with the single model,Catboost improves the generalization ability and prediction accuracy of the prediction method.Compared with other integrated learning frameworks that improve GBDT,Catboost effectively addresses the prediction bias caused by outliers through a new strategy of selecting tree structure.In this paper,the prediction performance of Cat Boost for short-term traffic flow prediction is further improved by integrating time series features.(2)Short-term traffic flow prediction model(NIKF)based on noise immune Kalman filtering: In this paper,the original data are decomposed and reconstructed by discrete wavelet transform to maintain the low frequency approximate part of the basic mode of the original traffic flow data and the high frequency detail part containing the abrupt details and noise together as the input of the model.A new cost function is used to reconstruct the traditional Kalman filter,and a novel and effective prediction model is proposed,which can obtain high quality and effective signals while efficient noise filtering,and the prediction results are further modified by using the periodic time-series correlation of traffic flow.